The generation of photo-realistic images is a major topic in computer graphics. By using the principles of physical light propagation, images that are indistinguishable from real photographs can be generated. However, this computation is a very time-consuming task. When simulating the real behavior of light, images can take hours to be of sufficient quality. This paper proposes a bio-inspired architecture with spiking neurons for fast rendering in global illumination. The objective is to find the number of paths that are required for each image in order to be perceived identical to the visually converged one computed by the path tracing algorithm. The challenge is that the visually converged image is unknown so that we start from a very noisy image to converge toward the less noisy image. This architecture with functional parts of sparse encoding, dynamic learning, and decoding consists of a robust convergence measure on blocks. Different pooling strategies are performed in order to separate noise from signal in a deep learning process. The learning algorithm selects the most pertinent images using clustering dynamic learning. The system dynamic computes a learning parameter for each image based on its level of noise. The experiments are conducted on a global illumination set which contains a large number of images with different resolutions and noise levels computed using diffuse and specular rendering. With respect to the scenes with 512×512 resolution, 3232 different images are used for learning and 9696 images are used for testing. For the scenes with 800×800 resolution, the training and the testing data contain, respectively, 3760 and 6320 images. The result is a system composed from only two spike pattern association neurons that accurately predict the quality of images with respect to human psycho-visual scores. The pooling spike neural network has been compared with the support vector and fast relevance vector machines. The obtained results show that the proposed method gives promising efficiency in terms of accuracy (which is calculated as the mean square error on each block of the scenes and the variation of the actual thresholds of the perception models and the desired human psycho-visual scores) and less number of parameters.